Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Learning for Diabetic Retinopathy Diagnosis: Enhancing Accuracy and Generalizability in Under-Resourced Regions

Gajan Mohan Raj, Michael G. Morley, Mohammad Eslami

TL;DR

The paper tackles diabetic retinopathy diagnosis in regions with limited ophthalmology resources by proposing a privacy-preserving federated learning framework using EfficientNetB0 to improve cross-institution generalizability. It leverages three public fundus datasets to simulate multi-institution training via FedAvg (TensorFlow Federated) and evaluates both accuracy on unseen data and generalization to degraded-image data. Results show the federated model surpasses local models (0.9321 vs 0.8876/0.9036/0.7148) and generalizes well across institutions, achieving 0.9708, 0.9634, and 0.9105 on H1/H2/H3 test sets when data quality varies. The work also demonstrates deployability with web and iOS prototypes, highlighting practical impact for under-resourced regions while outlining future work on privacy, scalability, and communication efficiency.

Abstract

Diabetic retinopathy is the leading cause of vision loss in working-age adults worldwide, yet under-resourced regions lack ophthalmologists. Current state-of-the-art deep learning systems struggle at these institutions due to limited generalizability. This paper explores a novel federated learning system for diabetic retinopathy diagnosis with the EfficientNetB0 architecture to leverage fundus data from multiple institutions to improve diagnostic generalizability at under-resourced hospitals while preserving patient-privacy. The federated model achieved 93.21% accuracy in five-category classification on an unseen dataset and 91.05% on lower-quality images from a simulated under-resourced institution. The model was deployed onto two apps for quick and accurate diagnosis.

Federated Learning for Diabetic Retinopathy Diagnosis: Enhancing Accuracy and Generalizability in Under-Resourced Regions

TL;DR

The paper tackles diabetic retinopathy diagnosis in regions with limited ophthalmology resources by proposing a privacy-preserving federated learning framework using EfficientNetB0 to improve cross-institution generalizability. It leverages three public fundus datasets to simulate multi-institution training via FedAvg (TensorFlow Federated) and evaluates both accuracy on unseen data and generalization to degraded-image data. Results show the federated model surpasses local models (0.9321 vs 0.8876/0.9036/0.7148) and generalizes well across institutions, achieving 0.9708, 0.9634, and 0.9105 on H1/H2/H3 test sets when data quality varies. The work also demonstrates deployability with web and iOS prototypes, highlighting practical impact for under-resourced regions while outlining future work on privacy, scalability, and communication efficiency.

Abstract

Diabetic retinopathy is the leading cause of vision loss in working-age adults worldwide, yet under-resourced regions lack ophthalmologists. Current state-of-the-art deep learning systems struggle at these institutions due to limited generalizability. This paper explores a novel federated learning system for diabetic retinopathy diagnosis with the EfficientNetB0 architecture to leverage fundus data from multiple institutions to improve diagnostic generalizability at under-resourced hospitals while preserving patient-privacy. The federated model achieved 93.21% accuracy in five-category classification on an unseen dataset and 91.05% on lower-quality images from a simulated under-resourced institution. The model was deployed onto two apps for quick and accurate diagnosis.

Paper Structure

This paper contains 21 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Motivations
  3. Ophthalmologists Shortage
  4. Current Challenges of AI in DR Diagnosis
  5. Resource Constraints in DR Imaging
  6. Solutions
  7. Materials and Methodology
  8. Background
  9. Federated learning-driven system framework
  10. Convolutional Neural Networks (CNNs)
  11. Dataset
  12. Implementation Details
  13. Preliminary Study (Finding Best CNN Architecture)
  14. Federation of Institutions' Diagnosis Models
  15. Prototype Demonstration
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Centralized data-sharing system (a) vs Federated learning-driven system (b)
  • Figure 2: FederatedAveraging Algorithm mcmahan_communication_efficient
  • Figure 3: Comparison of federated experiments: Experiment 1 (a) and Experiment 2 (b)
  • Figure 4: Developed demonstrations and applications.
  • Figure 5: Confusion Matrix of Federated Model on Independent Test-set
  • ...and 1 more figures